Multi-Model Splicer

Splicer sends your prompt to several models in parallel, then uses a judge to synthesize their complete answers. You can read each model's response, compare differing opinions, and see an agreement score. Agreement measures similarity between answers; it does not verify correctness.

How to run a splice

1. Open the circular navigation button on the right and choose Splicer.
2. Choose a preset or select at least two distinct models in Splicer settings. On desktop, settings appear to the left of the conversation; on narrow screens, the Splicer settings button opens a drawer.
3. Check the models and estimated cost, then send your prompt once.

Available models depend on your plan and current model availability. Splicer currently accepts text messages; use ordinary chat for file attachments.

The run shows how many model answers are complete. Its synthesized answer can include:

• Agreement % — how closely the complete answers align.
• Synthesized answer — the judge's combined response.
• Differing opinions — points where the models disagreed.

Read and keep the available answers

Under Model responses, choose a model tab or select Show all models. Responses are labeled Waiting, Responding, Complete, Partial, or Failed. A model that times out, reaches its output limit, or stops early can leave useful partial text. That text remains readable but does not count toward consensus.

Use Expand for a larger reading view, or Copy and Download to keep a synthesis or individual response. All run responses and details exports the available responses with run information.

Synthesis needs at least two complete answers. A run can synthesize with only some of your selected models; the result identifies partial coverage. If synthesis fails or the run is cancelled, read the available model responses below the status message.

Finish or stop a run

Once at least two model answers are complete, Finish with available answers lets you end the outstanding model requests and request synthesis from the complete answers. The run then shows Finishing with available answers or Synthesizing available answers.

To end a run without requesting a new synthesis, choose Stop run…, then confirm Stop without synthesis. Available answers and recorded charges are retained. A control request can take time to finish; check the run's final status rather than assuming the button response means settlement is complete.

Splicer also has per-model and overall time limits. After two complete answers, a short grace period allows remaining models to finish before synthesis proceeds. A slow or interrupted model does not have to prevent a synthesis from the complete answers.

Saved runs and interrupted connections

Reopen the conversation to load its saved Splicer run. Run details shows the run ID, outcome, and recorded reason when available. Use Reload saved run to check saved responses and billing status; this reads the existing run without starting another generation.

Refreshing or losing the connection can interrupt an active run. Keep any text already shown, reload the saved run, and check its outcome before sending the prompt again. A failed detail load keeps available text visible and offers a retry. Older runs may have less recorded detail.

Sources and context

Splicer shares the included conversation context, applicable project instructions, and retrieved evidence with the models and judge. The result reports available context information, including older messages omitted by the context limit and historical memory excerpts reused.

Check Live sources retrieved and the listed links when you need current information. Web restrictions and DynaRAG settings affect retrieval. Live sources not retrieved, or an unavailable source status on an older saved run, is not evidence that claims were checked against current sources. An agreement score alone cannot establish that a time-sensitive claim is current.

What it costs

The estimate in Splicer settings is a guide. Before dispatch, Nyquest reserves funds for the selected models and judge within their output allowances. Expensive models or long context can require a larger reservation than the eventual charge.

The run separates answer status from billing:

• Recorded cost or Known cost so far shows the available cost information.
• Charged shows the amount charged; Still reserved shows funds held while provider charges are reconciled.
• Billing and Cost status identify whether accounting is settled or still unresolved.

A partial, failed, or stopped answer can still incur provider charges. Settled billing does not mean the answer completed successfully. Confirmed unused funds are released on settlement; unknown provider costs keep their reservation until reconciled. Check the saved run and Usage before retrying an interrupted request.

Compression is enabled by default and may optimize each model's request. It can be bypassed when unavailable or unsuitable. Any displayed input-savings estimate is not a measured wallet saving or billing credit. API callers can opt out with splicer.compress: false.

Limits

The current request limit is six distinct models even on plans without a separate model-count cap. Each account can have up to two active Splicer runs. Wallet, model eligibility, and technical spending limits still apply.

Daily admission checks include charges and outstanding reservations. The daily window resets at midnight UTC, but unresolved holds from earlier dates continue to count until reconciled. A new run can be rejected before dispatch if its reservation will not fit the remaining allowance.

Developer API

Submit a text run through POST /v1/chat/completions with x-nyquest-mode: splicer or mode: "splicer" and a splicer configuration. These endpoints inspect and control your own runs:

Finish and stop acknowledge a control request; use the stream or run-detail endpoint to check the final outcome. They do not create a replacement run. See the Endpoints Reference for the wider API.

Where to next

• Streaming — recover an interrupted ordinary-chat answer.
• Auto-Router: Tiers, Domains, and Providers — how single-model routing works.
• Wallet and Funds — funding requests and reviewing usage.